01
Expand
The query is rewritten into alternative phrasings. Different wordings pull back different documents, which widens recall before ranking narrows it again.
From $2 per 1,000 searches, full page content included.
Search engines rank pages for people to click. Agents do not click, they read. One call fans out across engines, fuses the rankings, opens the pages, ranks the passages inside them and returns an answer whose every quote is checked back against its source.
One credit balance across every endpoint. A basic search is one credit, a deep search that opens pages is two, a grounded answer adds one. Credits do not expire, there is no monthly fee and no per seat charge.
| Search API | Cost per 1,000 searches | What that buys |
|---|---|---|
| Neurobird Search | $2 basic, $4 deep, $6 deep with answer | Ranked results, full page content, grounded answer with verified quotes |
| Serper | $1.00, down to $0.30 at 12.5M | Google SERP links only, no page content |
| Parallel Search | from $1 | Ranked URLs with compressed excerpts |
| Brave Search API | $5 | Independent index, results only |
| Google Custom Search | $5 | Results only, hard cap of 10k queries a day |
| Exa | about $7 | Neural search, page contents billed separately |
| Tavily | $8 basic, $16 advanced | Search with extraction, credit metered |
| Linkup | from $10 | Search, minimum charge per request |
| Anthropic web search tool | $10 | Search inside the Claude API, plus token costs |
| Grounding with Bing (Azure) | $35 | Bing results for grounding |
Competitor figures were read from each vendor's own pricing or docs page in August 2026 and converted to a per 1,000 basis. Tavily bills $0.008 per credit with an advanced search costing two credits. Serper's rate depends on pack size. Vendors change prices, so treat this as the date stamped snapshot it is rather than a live feed.
A SERP wrapper hands back ten links and leaves the hard part to you. Every stage below runs inside one request, under a wall clock budget, so a slow engine costs one source instead of the whole response.
The query is rewritten into alternative phrasings. Different wordings pull back different documents, which widens recall before ranking narrows it again.
Every query hits a self hosted metasearch layer that queries many engines at once and reports which engine found each result, and where it placed.
Reciprocal rank fusion merges those independent rankings. Engine scores are not comparable, but agreement between engines is the strongest cheap relevance signal there is.
Top candidates are fetched through a pool of unmetered ISP addresses and stripped to main content. Browser grade TLS fingerprints remove the false 403s that look like dead links.
Pages are split into passages and scored by phrase aware BM25 fused with embedding similarity. You get the part of the page that answers the question, not the page.
The answer is written only from retrieved passages, and each supporting quote is matched back against the text it came from. Quotes that do not match are returned marked unverified.
Every citation carries a quote, and every quote is matched against the page it is attributed to before the response leaves. Unverified quotes are labelled, not hidden. A summary you cannot audit is a summary you cannot ship.
Send a JSON Schema with the query and get that schema filled from each result, in the same call. Fields the page does not state come back null rather than invented.
Ranked links arrive over server sent events as soon as fusion completes, seconds before the summary exists. An agent can start its next tool call while the answer is still being written.
Each response reports how long expansion, retrieval, reading, ranking and answering took. Latency here is set by whichever upstream is slowest that minute, and you should be able to see that from the outside.
Egress runs over unmetered addresses, so returning complete page text costs no extra credit. Extraction is not an upsell here, it is the default the pipeline already needed.
The Model Context Protocol endpoint is hosted here over streamable HTTP. One command adds it to Claude Code, and the same URL works in Cursor, Windsurf, Zed and the Claude desktop app.
No SDK required. One POST with a JSON body, so it works from a terminal, from any language, and from an agent framework without a plugin.
One click issues a key with 1,000 credits a month. No email, no card, no onboarding call. If it does not do what you need, you have lost a click.
Neurobird Search costs $2 per 1,000 basic searches, $4 per 1,000 deep searches and $6 per 1,000 deep searches with a grounded answer, dropping to $1 per 1,000 credits above one million credits a month. Full page content is included at no extra credit. For comparison, on prices published by each vendor in August 2026: Tavily bills $0.008 per credit, which is $8 per 1,000 basic searches and $16 per 1,000 advanced; Exa lists $7 per 1,000 search requests; Brave Search API lists $5 per 1,000 requests; Serper lists $1.00 per 1,000 on a 50,000 credit pack, falling to $0.30 per 1,000 on a 12.5 million credit pack; the Anthropic web search tool is $10 per 1,000 searches plus token costs.
Yes. Neurobird Search accepts the same request shape and returns the same response shape as Tavily, so an existing client switches by changing the base URL to https://search.neurobird.com and the bearer token to a Neurobird key. Basic search is $2 per 1,000 against Tavily's $8 per 1,000, and full page content is returned without a separate extract call.
Neurobird Search issues 1,000 credits every month on a key created by a single unauthenticated POST to https://search.neurobird.com/keys. There is no signup form, no email address and no card. Every endpoint and every feature is available on the free tier, including grounded answers and quote verification.
Neurobird Search exposes an MCP server over streamable HTTP at https://search.neurobird.com/mcp with two tools, web_search and extract_url. There is nothing to install or run locally. Add it to Claude Code with: claude mcp add --transport http neurobird-search https://search.neurobird.com/mcp --header "Authorization: Bearer $NEUROBIRD_KEY".
Neurobird Search fetches and extracts the main content of each ranked result and returns it in the same response, at no additional credit. SERP APIs such as Serper and Google Custom Search return links and snippets only. Exa bills page contents separately at $1 per 1,000 pages on its published price list.
The answer model is required to return verbatim quotes, and each quote is matched back against the extracted source text before the response is returned. Whitespace and quote glyphs are normalised. A quote that still fails the match is returned marked verified: false rather than silently dropped, so a caller can audit every claim.
Measured medians on the current single-server deployment are about 15 seconds for a basic search with an answer, about 22 seconds at standard depth and about 28 seconds at advanced depth. Upstream engines and the answer model account for nearly all of that. With the streaming endpoint, ranked links reach the client at about 5 seconds and the answer begins writing at about 11 seconds. Every response carries per stage timings so latency can be attributed from outside.